One hour
$3.59
1 GPU × 1h · compute only
High-memory Hopper GPU
Rent an NVIDIA H200 GPU for memory-bound AI.
Direct answer
NVIDIA H200 pairs Hopper compute with 141 GB of HBM3e, making it useful when model weights, KV cache or scientific data exceed an 80 GB device. GPURento lists a $3.59 per GPU-hour on-demand rate with self-service creation in Paris and Frankfurt.
Reviewed by the GPURento infrastructure team · September 1, 2026
Cost scenarios
What one H200 SXM costs at the listed rate.
One day
$86.16
1 GPU × 24h · compute only
Seven days
$603.12
1 GPU × 168h · compute only
Average month · 730h
$2620.70
1 GPU × 730h · compute only
| Memory | 141 GB HBM3e | More single-GPU memory than H100 SXM. |
|---|---|---|
| Memory bandwidth | 4.8 TB/s | Manufacturer specification for H200. |
| On-demand rate | $3.59 / GPU-hour | Compute only, before storage or optional services. |
| Available regions | Paris · Frankfurt | Both regions are selectable after funding. |
Best for
- Large-model inference with a substantial KV cache
- Models that narrowly exceed 80 GB
- Memory-bound HPC and data analytics
- Reducing tensor or pipeline parallelism for some workloads
Not the best fit when
- Jobs that fit comfortably on 24–48 GB
- Price-first development notebooks
- Compute-bound jobs that do not benefit from extra memory
Decision method
What to verify before you deploy capacity.
The content below separates published specifications, GPURento catalog references and decisions that still require a workload benchmark.
01
The main H200 advantage is memory
H200 is not simply “a faster H100.” Its practical advantage is the larger, higher-bandwidth memory subsystem. That can let a model use fewer GPUs, hold a longer context or serve more concurrent sequences, depending on quantization and runtime.
- Calculate weights and runtime overhead
- Reserve KV-cache headroom for target context and concurrency
- Test whether fewer GPUs offset the higher hourly rate
02
Compare cost per request, not only cost per hour
For inference, record time to first token, output tokens per second, p95 latency and cost per million tokens at a fixed quality level. A high-memory GPU is valuable only if the runtime can use that memory and bandwidth efficiently.
03
Two EU deployment regions
The catalog exposes H200 in EU West Paris and EU Central Frankfurt. GPURento checks account funding, stores the configuration and starts the self-service provisioning workflow without a sales form.
Catalog shortlist
Relevant GPU options.
HopperAvailable
NVIDIA H200 SXM
141 GB HBM3e · $3.59/hr
Large-model inference and HPC
EU West · ParisEU Central · Frankfurt
HopperAvailable
NVIDIA H100 SXM
80 GB HBM3 · $2.69/hr
Intensive training and FP8 inference
EU West · ParisEU Central · Frankfurt
BlackwellAvailable
NVIDIA B200
180 GB HBM3e · $5.98/hr
Frontier training and inference
EU West · ParisEU Central · Frankfurt
Questions
Clear answers, including the limits.
How much memory does an H200 have?+
The H200 configuration described here has 141 GB of HBM3e memory and a manufacturer-specified 4.8 TB/s memory bandwidth.
When should I choose H200 instead of H100?+
Choose H200 when 80 GB forces unwanted sharding, limits context or constrains concurrency. If your workload already fits and is compute-bound, benchmark both before paying the higher reference rate.
Where can I deploy H200 capacity?+
The current GPURento catalog exposes self-service H200 creation in EU West Paris and EU Central Frankfurt.
Sources and scope
Last reviewed September 1, 2026. GPURento rates are current catalog references; provisioning state remains visible in the workspace, and manufacturer specifications do not substitute for workload benchmarks.
Continue the research
Deployment plan
Check whether H200 removes a memory bottleneck.
Create a workspace and choose Paris or Frankfurt with your model size, runtime and target concurrency.